bayesian inference for the calibration of dsmc parameters
DESCRIPTION
Bayesian Inference For The Calibration Of DSMC Parameters. James S. Strand and David B. Goldstein The University of Texas at Austin. Sponsored by the Department of Energy through the PSAAP Program. Computational Fluid Physics Laboratory. Predictive Engineering and Computational Sciences. - PowerPoint PPT PresentationTRANSCRIPT
![Page 1: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/1.jpg)
Bayesian Inference For The Calibration Of DSMC Parameters
James S. Strand and David B. GoldsteinThe University of Texas at Austin
Sponsored by the Department of Energy through the PSAAP Program
Predictive Engineering and Computational Sciences
Computational Fluid Physics Laboratory
![Page 2: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/2.jpg)
Motivation – DSMC Parameters• The DSMC model includes many parameters related to gas dynamics at the molecular level, such as: Elastic collision cross-sections. Vibrational and rotational excitation probabilities. Reaction cross-sections. Sticking coefficients and catalytic efficiencies for gas-
surface interactions. …etc.
![Page 3: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/3.jpg)
DSMC Parameters
• In many cases the precise values of some of these parameters are not known.• Parameter values often cannot be directly measured, instead they must be inferred from experimental results.• By necessity, parameters must often be used in regimes far from where their values were determined.• More precise values for important parameters would lead to better simulation of the physics, and thus to better predictive capability for the DSMC method. The ultimate goal of this work is to use experimental data to calibrate important DSMC parameters.
![Page 4: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/4.jpg)
DSMC Method
Direct Simulation Monte Carlo (DSMC) is a particle based simulation method.
• Simulated particles represent large numbers of real particles.• Particles move and undergo collisions with other particles.• Can be used in highly non-equilibrium flowfields (such as strong shock waves). • Can model thermochemistry on a more detailed level than most CFD codes.
![Page 5: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/5.jpg)
DSMC Method
Initialize
Move
Index
Collide
Sample
Performed on First Time Step
Performed on Selected Time Steps
Performed on Every Time Step
Create
![Page 6: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/6.jpg)
DSMC Collisions
DSMC collisions are performed statistically. Pairs of molecules are randomly selected from within a cell, a collision probability is calculated for each selected pair, and a random number draw determines whether or not the pair actually collides.
![Page 7: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/7.jpg)
DSMC Collisions
DSMC collisions are performed statistically. Pairs of molecules are randomly selected from within a cell, a collision probability is calculated for each selected pair, and a random number draw determines whether or not the pair actually collides.
![Page 8: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/8.jpg)
DSMC Collisions
DSMC collisions are performed statistically. Pairs of molecules are randomly selected from within a cell, a collision probability is calculated for each selected pair, and a random number draw determines whether or not the pair actually collides.
![Page 9: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/9.jpg)
Selection Possibilities
![Page 10: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/10.jpg)
DSMC Collisions
If the pair does not collide, their properties are left unchanged and a new pair is selected. If the pair does collide, the properties of both colliding particles are immediately adjusted to their calculated post-collision values, and then a new pair is selected. The number of selections in a given cell is a function of the overall simulation conditions and of the conditions within that cell at that time step.
![Page 11: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/11.jpg)
DSMC Collisions
If the pair does not collide, their properties are left unchanged and a new pair is selected. If the pair does collide, the properties of both colliding particles are immediately adjusted to their calculated post-collision values, and then a new pair is selected. The number of selections in a given cell is a function of the overall simulation conditions and of the conditions within that cell at that time step.
![Page 12: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/12.jpg)
Variable Hard Sphere ModelThe VHS model allows the collision cross-section to be dependent on relative speed.
There are two relevant parameters for the VHS model, dref and ω. Both of these parameters are usually determined based on viscosity data.
𝝈𝑽𝑯𝑺=𝝅𝒅𝑽𝑯𝑺=𝝅𝒅𝒓𝒆𝒇 (𝒄𝒓 ,𝒓𝒆𝒇𝒄𝒓 )(𝝎−𝟏𝟐 )
![Page 13: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/13.jpg)
Numerical Methods – DSMC Code
• Our DSMC code can model flows with rotational and vibrational excitation and relaxation, as well as five-species air chemistry, including dissociation, exchange, and recombination reactions.• Larsen-Borgnakke model is used for redistribution between rotational, translational, and vibrational modes during inelastic collisions.• TCE model allows cross-sections for chemical reactions to be derived from Arrhenius parameters.
![Page 14: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/14.jpg)
Internal Modes
• Rotation is assumed to be fully excited. Each particle has its own value of rotational energy,
and this variable is continuously distributed.• Vibrational levels are quantized.
Each particle has its own vibrational level, which is associated with a certain vibrational energy based on the simple harmonic oscillator model.
• Relevant parameters are ZR and ZV, the rotational and vibrational collision numbers.
ZR = 1/ΛR, where ΛR is the probability of the rotational energy of a given molecule being redistributed during a given collision.
ZV = 1/ΛV
![Page 15: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/15.jpg)
Chemistry Implementation
Reaction cross-sections based on Arrhenius rates TCE model allows determination of reaction cross-
sections from Arrhenius parameters.
, the average number of internal degrees of freedom which contribute to the collision energy.
is the temperature-viscosity exponent for VHS collisions between type A and type B particles
𝜎 𝑟𝑒𝑓∧𝑇 𝑟𝑒𝑓 are both constants related ¿ the VHS collisionmodel𝜀=1 (𝑖𝑓 𝐴≠𝐵 )𝑜𝑟 2 (𝑖𝑓 𝐴=𝐵 )
σR and σT are the reaction and total cross-sections, respectively
k is the Boltzmann constant, mr is the reduced mass of particles A and B, Ec is the collision energy, and Γ() is the gamma function.
![Page 16: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/16.jpg)
Reactions𝑘 (𝑇 )=𝚲𝑇𝜼𝑒−𝑬𝒂 /𝑘𝑇
R. Gupta, J. Yos, and R. Thompson, NASA Technical Memorandum 101528, 1989.
# Reaction Forward Rate Constants Backward Rate Constants
qreaction Λ η EA Λ η EA
1 N2 + N2 <--> N2 + N + N 7.968E-13 -0.5 1.561E-18 6.518E-47 0.27 0.0 -1.561E-18 2 N + N2 <--> N + N + N 6.9E-8 -1.5 1.561E-18 4.817E-46 0.27 0.0 -1.561E-18 3 O2 + N2 <--> O2 + N + N 3.187E-13 -0.5 1.561E-18 6.518E-47 0.27 0.0 -1.561E-18 4 O + N2 <--> O + N + N 3.187E-13 -0.5 1.561E-18 6.518E-47 0.27 0.0 -1.561E-18 5 NO + N2 <--> NO + N + N 3.187E-13 -0.5 1.561E-18 6.518E-47 0.27 0.0 -1.561E-18 6 N2 + O2 <--> N2 + O + O 1.198E-11 -1.0 8.197E-19 1.8E-47 0.27 0.0 -8.197E-19 7 N + O2 <--> N + O + O 5.993E-12 -1.0 8.197E-19 1.8E-47 0.27 0.0 -8.197E-19 8 O2 + O2 <--> O2 + O + O 5.393E-11 -1.0 8.197E-19 1.8E-47 0.27 0.0 -8.197E-19 9 O + O2 <--> O + O + O 1.498E-10 -1.0 8.197E-19 1.8E-47 0.27 0.0 -8.197E-19 10 NO + O2 <--> NO + O + O 5.993E-12 -1.0 8.197E-19 1.8E-47 0.27 0.0 -8.197E-19 11 N2 + NO <--> N2 + N + O 6.59E-10 -1.5 1.043E-18 2.0976E-46 0.27 0.0 -1.043E-18 12 N + NO <--> N + N + O 1.318E-8 -1.5 1.043E-18 2.0976E-46 0.27 0.0 -1.043E-18 13 O2 + NO <--> O2 + N + O 6.59E-10 -1.5 1.043E-18 2.0976E-46 0.27 0.0 -1.043E-18 14 O + NO <--> O + N + O 1.318E-8 -1.5 1.043E-18 2.0976E-46 0.27 0.0 -1.043E-18 15 NO + NO <--> NO + N + O 1.318E-8 -1.5 1.043E-18 2.0976E-46 0.27 0.0 -1.043E-18 16 N2 + O <--> NO + N 1.12E-16 0.0 5.175E-19 2.490E-17 0.0 0.0 -5.175E-19 17 NO + O <--> O2 + N 5.279E-21 1.0 2.719E-19 1.598E-18 0.5 4.968E-20 -2.719E-19
![Page 17: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/17.jpg)
Collision Rates
Temperature (K)
CollisionRate(#/m
3 )
5000 10000 15000 20000 25000
1E+29
3E+29
5E+29
7E+29
N2 - N2 (DSMC)N2 - N2 (VHS)N2 - N (DSMC)N2 - N (VHS)N - N (DSMC)N - N (VHS)N - O (DMSC)N - O (VHS)O2 - NO (DSMC)O2 - NO (VHS)
VHS Collision Pairs
![Page 18: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/18.jpg)
Reaction Rates – Dissociation
Temperature (K)
ReactionRate(#/m
3 )
5000 10000 15000 20000 25000
1.0E+28
3.0E+28
5.0E+28
7.0E+28N2 + N2 --> N + N + N2 (DSMC)N2 + N2 --> N + N + N2 (Arrhenius)N2 + N --> N + N + N (DSMC)N2 + N --> N + N + N (Arrhenius)O2 + N2 --> O + O + N2 (DSMC)O2 + N2 --> O + O + N2 (Arrhenius)NO + O --> N + O + O (DSMC)NO + O --> N + O + O (Arrhenius)
Dissociation Reactions
![Page 19: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/19.jpg)
Temperature (K)
ReactionRate(#/m
3 )
5000 10000 15000 20000 25000
1E+28
3E+28
5E+28
N + N + N2 --> N2 + N2 (DSMC)N + N + N2 --> N2 + N2 (Arrhenius)N + N + N --> N2 + N (DSMC)N + N + N --> N2 + N (Arrhenius)O + O + O --> O2 + O (DSMC)O + O + O --> O2 + O (Arrhenius)N + O + N --> NO + N (DSMC)N + O + N --> NO + N (Arrhenius)
Recombination Reactions
Reaction Rates – Recombination
![Page 20: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/20.jpg)
Reaction Rates – Exchange
Temperature (K)
ReactionRate(#/m
3 )
5000 10000 15000 20000 25000
2E+28
4E+28
6E+28
8E+28
N2 + O --> NO + N (DSMC)N2 + O --> NO + N (Arrhenius)O2 + N --> NO + O (DSMC)O2 + N --> NO + O (Arrhenius)NO + N --> N2 + O (DSMC)NO + N --> N2 + O (Arrhenius)NO + O --> O2 + N (DSMC)NO + O --> O2 + N (Arrhenius)
Exchange Reactions
![Page 21: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/21.jpg)
Parallelization
• DSMC: MPI parallel. Ensemble averaging to reduce stochastic noise for
0-D relaxations. Adaptive load balancing for 1-D shock simulations.
• Sensitivity Analysis and MCMC: MPI Parallel Separate processor groups call DSMC subroutine
independently.
![Page 22: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/22.jpg)
Sensitivity Analysis - Overview
• In the current context, the goal of sensitivity analysis is to determine which parameters most strongly affect a given quantity of interest (QoI). • Only parameters to which a given QoI is sensitive will be informed by calibrations based on data for that QoI.
![Page 23: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/23.jpg)
Sensitivity Analysis Parameters𝑘 (𝑇 )=𝚲𝑇𝜼𝑒−𝑬𝒂 /𝑘𝑇 10𝛼=𝚲
Throughout the sensitivity analysis, the ratio of forward to backward rate for a given reaction is kept constant, since these ratios should be fixed by the equilibrium constant.
# Reaction αmin αnom αmax Nominal Forward Rate Constants
Λ η EA 1 N2 + N2 <--> N2 + N + N -13.099 -12.099 -11.099 7.968E-13 -0.5 1.561E-18 2 N + N2 <--> N + N + N -8.161 -7.161 -6.161 6.9E-8 -1.5 1.561E-18 3 O2 + N2 <--> O2 + N + N -13.497 -12.497 -11.497 3.187E-13 -0.5 1.561E-18 4 O + N2 <--> O + N + N -13.497 -12.497 -11.497 3.187E-13 -0.5 1.561E-18 5 NO + N2 <--> NO + N + N -13.497 -12.497 -11.497 3.187E-13 -0.5 1.561E-18 6 N2 + O2 <--> N2 + O + O -11.922 -10.922 -9.922 1.198E-11 -1.0 8.197E-19 7 N + O2 <--> N + O + O -12.222 -11.222 -10.222 5.993E-12 -1.0 8.197E-19 8 O2 + O2 <--> O2 + O + O -11.268 -10.268 -9.268 5.393E-11 -1.0 8.197E-19 9 O + O2 <--> O + O + O -10.824 -9.824 -8.824 1.498E-10 -1.0 8.197E-19 10 NO + O2 <--> NO + O + O -12.222 -11.222 -10.222 5.993E-12 -1.0 8.197E-19 11 N2 + NO <--> N2 + N + O -10.181 -9.181 -8.181 6.59E-10 -1.5 1.043E-18 12 N + NO <--> N + N + O -8.880 -7.880 -6.880 1.318E-8 -1.5 1.043E-18 13 O2 + NO <--> O2 + N + O -10.181 -9.181 -8.181 6.59E-10 -1.5 1.043E-18 14 O + NO <--> O + N + O -8.880 -7.880 -6.880 1.318E-8 -1.5 1.043E-18 15 NO + NO <--> NO + N + O -8.880 -7.880 -6.880 1.318E-8 -1.5 1.043E-18 16 N2 + O <--> NO + N -16.951 -15.951 -14.951 1.12E-16 0.0 5.175E-19 17 NO + O <--> O2 + N -21.277 -20.277 -19.277 5.279E-21 1.0 2.719E-19
![Page 24: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/24.jpg)
Scenario: 1-D Shock
• Shock speed is ~8000 m/s, M∞ ≈ 23.• Upstream number density = 3.22×1021 #/m3.• Upstream composition by volume: 79% N2, 21% O2.• Upstream temperature = 300 K.
![Page 25: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/25.jpg)
1D Shock Simulation
X
Pressure
StreamwiseVelocity
InletBoundary SpecularW
all
Simulation is initialized with a bulk velocitydirected towards the specular wall at the rightboundary of the domain, with pre-shock density,temperature, and chemical composition.
Vbulk
X
Pressure
StreamwiseVelocity
InletBoundary SpecularW
all
ShockPropagatesUpstream
X
Pressure
StreamwiseVelocity
InletBoundary SpecularW
all
The shock is allowedto move a reasonabledistance away from thewall before any form ofsampling begins.
X
Upstream PressureSampling Region
InletBoundary
SpecularWallPressure
Downstream Pressure Sampling Region
X0.0
0.5
1.0
1.5 Normalized PressureBoxcar Averaged Normalized Pressure
InletBoundary
SpecularWall
Shock Position
X0.0
0.5
1.0
1.5 Boxcar Averaged Normalized Pressure (Time = t)Boxcar Averaged Normalized Pressure (Time = t + t)
InletBoundary
SpecularWall
Shock Position(time = t)
Shock Position(time = t + t)
sShock PropagationSpeed = VP = s/t
X0.0
0.5
1.0
1.5 Boxcar Averaged Normalized Pressure (Time = t + t)
InletBoundary
SpecularWallShock Position
(time = t + t)
Shock SamplingRegion
VSampling Region = VP
• We require the ability to simulate a 1-D shock without knowing the post-shock conditions a priori.• To this end, we simulate an unsteady 1-D shock, and make use of a sampling region which moves with the shock.
![Page 26: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/26.jpg)
Results: Nominal Parameter Values
X (m)
Density(kg/m
3 )
-0.05 0 0.05 0.1 0.15 0.20.0000
0.0005
0.0010
0.0015
0.0020
0.0025
0.0030 BulkN2NO2ONO
Bulk
O2, NO
N
O
N2
X (m)
Temperature(K)
0 0.05 0.1 0.15 0.20
5000
10000
15000
20000
25000
TtransTrotTvibT
X (m)
N 2Temperature(K)
-0.01 0 0.01 0.020
5000
10000
15000
20000
25000TtransTrotTvibT
![Page 27: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/27.jpg)
Quantity of Interest (QoI)
J. Grinstead, M. Wilder, J. Olejniczak, D. Bogdanoff, G. Allen, and K. Danf, AIAA Paper 2008-1244, 2008.
X
QoI
We cannot yet simulate NASA EAST or other shock tube results, so we must choose a temporary, surrogate QoI.
?
![Page 28: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/28.jpg)
{𝑸𝒐𝑰 }={𝑸𝒐𝑰𝟏𝑸𝒐𝑰𝟐𝑸𝒐𝑰𝟑
⋮𝑸𝒐𝑰𝒏
}
Sensitivity Analysis - Scalar vs. Vector QoI
The density of NO (our QoI for this work) is a vector, with values over a range of points in space.
X (m)
NO(kg/m
3 )
0 0.05 0.1 0.15 0.20.0E+00
1.0E-05
2.0E-05
3.0E-05
4.0E-05
5.0E-05
6.0E-05
7.0E-05
For our analysis, each blue dotrepresents a single, scalar QoI.
Scalar QoI used in later schematicsdemonstrating calculation of correlationcoefficients and the mutual information.
![Page 29: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/29.jpg)
Sensitivity Analysis - Methods
Two measures for sensitivity were used in this work.• Pearson correlation coefficients:
• The mutual information:
Both measures involve global sensitivity analysis based on a Monte Carlo sampling of the parameter space, and thus the same datasets can beused to obtain both measures.
![Page 30: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/30.jpg)
log10(O2 + NO <--> O + O + NO)
NO(kg/m
3 ),atx=-0.00016m
-12 -11.5 -11 -10.50
2E-05
4E-05
6E-05 r = -0.0346r2 = 0.0012
log10(NO + N <--> N + O + N)
NO(kg/m
3 ),atx=-0.00016m
-8.5 -8 -7.5 -70
2E-05
4E-05
6E-05
r = -0.3006r2 = 0.0904
log10(O2 + N <--> NO + O)
NO(kg/m
3 ),atx=-0.00016m
-18.5 -18 -17.5 -170
2E-05
4E-05
6E-05
r = 0.6045r2 = 0.3655
Sensitivity Analysis: Correlation Coefficient
![Page 31: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/31.jpg)
1
QoI
1
QoI
Sensitivity Analysis: Mutual Information
![Page 32: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/32.jpg)
Sensitivity Analysis: Mutual Information
1
p( 1)
QoI
p(QoI)
0.2460.1850.1230.0620.000
p(1,QoI)
![Page 33: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/33.jpg)
Sensitivity Analysis: Mutual Information
1
p( 1)
QoI
p(QoI)
0.1150.0860.0570.0290.000
p(1)p(QoI)
![Page 34: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/34.jpg)
Sensitivity Analysis: Mutual Information
𝑰 (𝜽𝟏 ,𝑸𝒐𝑰 )=∫𝜽𝟏
∫𝑸𝒐𝑰𝒑 (𝜽𝟏 ,𝑸𝒐𝑰 )[𝒍𝒏( 𝒑 (𝜽𝟏 ,𝑸𝒐𝑰 )
𝒑 (𝜽𝟏)𝒑 (𝑸𝒐𝑰 )) ]𝒅𝑸𝒐𝑰 𝒅𝜽𝟏
Hypothetical joint PDF for case where the QoI is indepenent of θ1.
Actual joint PDF of θ1 and the QoI, from a Monte Carlo sampling of theparameter space.
Kullback-Leibler divergence0.2460.1850.1230.0620.000
p(1,QoI)
0.1150.0860.0570.0290.000
p(1)p(QoI)0.0080.0060.0040.0020.000
𝐩 (𝛉𝟏 ,𝐐𝐨𝐈 )[𝐥𝐧 ( 𝐩 (𝛉𝟏 ,𝐐𝐨𝐈)𝐩 (𝛉𝟏 )𝐩(𝐐𝐨𝐈)) ]
Hypothetical joint PDF for case where the QoI is indepenent of θ1.
Actual joint PDF of θ1 and the QoI, from a Monte Carlo sampling of theparameter space.
QoI
θ1
![Page 35: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/35.jpg)
X (m)
NO(kg/m
3 )
0 0.05 0.1 0.15 0.20.0E+00
1.0E-05
2.0E-05
3.0E-05
4.0E-05
5.0E-05
6.0E-05
7.0E-05
For our analysis, each blue dotrepresents a single, scalar QoI.
Scalar QoI used in later schematicsdemonstrating calculation of correlationcoefficients and the mutual information.
X (m)
r2
MutualInformation
0 0.05 0.1 0.150
0.2
0.4
0.6
0.8
0
0.2
0.4
0.6
0.8
1N2 + N --> N + N + N (r2)
NO + N --> N + O + N (r2)NO + O --> N + O + N (r2)N2 + O --> NO + N (r
2)N2 + N --> N + N + N (MI)NO + N --> N + O + N (MI)NO + O --> N + O + N (MI)N2 + O --> NO + N (MI)
X (m)
r2
MutualInformation
0 0.05 0.1 0.150
0.2
0.4
0.6
0.8
0
0.2
0.4
0.6
0.8
1N2 + O --> NO + N (r2)
N2 + O --> NO + N (MI)
At this x-location, r2 is zerowhile the mutual informationis greater than zero.
Sensitivities vs. X
![Page 36: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/36.jpg)
r2 vs. Mutual Information
log10(N2 + O --> NO + N)
NO(kg/m
3 ),atx=0.021m
-16.5 -16.0 -15.5 -15.0
5.0E-06
1.0E-05
1.5E-05
2.0E-05
2.5E-05r2 = 0.0000002MI = 0.0668
X (m)
NO(kg/m
3 )
-0.01 0 0.01 0.02 0.03 0.04 0.050
2E-05
4E-05
6E-05
8E-05
Lowest for N2 + O <--> NO + NNominal for N2 + O <--> NO + NHighest for N2 + O <--> NO + N
![Page 37: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/37.jpg)
Sensitivities vs. X for ρNO as QoI
r2
MutualInformation
0 0.05 0.1 0.150
0.2
0.4
0.6
0.8
0
0.2
0.4
0.6
0.8
1N2 + N --> N + N + N (r2)
NO + N --> N + O + N (r2)N2 + N --> N + N + N (MI)NO + N --> N + O + N (MI)
X (m)
log10(N2 + N <--> N + N + N)
NO(kg/m
3 ),atx=0.141m
-8 -7.5 -7 -6.50.0E+00
5.0E-05
1.0E-04
1.5E-04
2.0E-04 r2 = 0.3053
log10(NO + N <--> N + O + N)
NO(kg/m
3 ),atx=0.003m
-8.5 -8 -7.5 -70.0E+00
5.0E-05
1.0E-04
1.5E-04
2.0E-04 r2 = 0.2902
![Page 38: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/38.jpg)
Variance of ρNO vs. X
![Page 39: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/39.jpg)
X (m)
var(
NO)xr2(Normalized)
var( N
O)xMI(Normalized)
0 0.05 0.1 0.150
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
N2 + N <--> N + N + N (r2)
NO + N <--> N + O + N (r2)NO + O <--> N + O + N (r2)N2 + O <--> NO + N (r
2)NO + O --> O2 + N (r
2)N2 + N <--> N + N + N (MI)NO + N <--> N + O + N (MI)NO + O <--> N + O + N (MI)N2 + O <--> NO + N (MI)NO + O --> O2 + N (MI)
X (m)
var(
NO)xr2(Normalized)
var( N
O)xMI(Normalized)
0 0.005 0.010
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1N2 + N <--> N + N + N (r2)
NO + N <--> N + O + N (r2)NO + O <--> N + O + N (r2)N2 + O <--> NO + N (r
2)NO + O --> O2 + N (r
2)N2 + N <--> N + N + N (MI)NO + N <--> N + O + N (MI)NO + O <--> N + O + N (MI)N2 + O <--> NO + N (MI)NO + O --> O2 + N (MI)
Variance Weighted Sensitivities
![Page 40: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/40.jpg)
Overall Sensitivities
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
Ove
rall
Sens
itivi
ty (N
orm
alize
d)
Parameter #
r^2 (QoI_2)
MI (QoI_2)
N2 + N <--> N + N + N
r2
Mutual Information
NO + N <--> N + O + N
NO + O <--> N + O + O
N2 + O <--> NO + N
NO + O <--> O2 + N
O2 + O <--> O + O + O
![Page 41: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/41.jpg)
Synthetic Data Calibrations: 0-D Relaxation
• Due to computational constraints, a 0-D relaxation from an initial high-temperature state was used for the synthetic data calibrations performed thus far.• 0-D box is initialized with 79% N2, 21% O2.
Initial bulk number density = 1.0×1023 #/m3. Initial bulk translational temperature = ~50,000 K. Initial bulk rotational and vibrational temperatures are
both 300 K.• Scenario is a 0-D substitute for a hypersonic shock at ~8 km/s.
Assumption that the translational modes equilibrate much faster than the internal modes.
• Sensitivity analysis of the type discussed earlier was used to identify parameters for calibration.
![Page 42: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/42.jpg)
Time (s)
NO(kg/m
3 )
0.0E+00 5.0E-07 1.0E-06 1.5E-06 2.0E-060.0E+00
5.0E-05
1.0E-04
1.5E-04
2.0E-04
2.5E-04
3.0E-04
Synthetic Data
Synthetic Data
• We once again use ρNO as our QoI.
![Page 43: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/43.jpg)
MCMC Method - Overview
• Markov Chain Monte Carlo (MCMC) is a method which solves the statistical inverse problem in order to calibrate parameters with respect to a set of data.• The likelihood of a given set of parameters is calculated based on the mismatch between the data and the simulation results for that set of parameters.• One or more chains explore the parameter space, moving towards regions of higher likelihood.• Candidate positions are drawn from a multi-dimensional Gaussian proposal distribution centered at the current chain position. The covariance matrix of this Gaussian controls the average distance (in parameter space) that the chain moves in one step.
![Page 44: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/44.jpg)
Time (s)
NO(kg/m
3 )
0.0E+00 5.0E-07 1.0E-06 1.5E-06 2.0E-060.0E+00
5.0E-05
1.0E-04
1.5E-04
2.0E-04
2.5E-04
3.0E-04
Synthetic Data
Time (s)
NO(kg/m
3 )
0.0E+00 5.0E-07 1.0E-06 1.5E-06 2.0E-060.0E+00
5.0E-05
1.0E-04
1.5E-04
2.0E-04
2.5E-04
3.0E-04
Synthetic Data2 Error Bars
𝒍𝒊𝒌𝒆𝒍𝒊𝒉𝒐𝒐𝒅=𝑷 (𝑫|𝜽 )= 𝟏
(𝟐𝝅𝝈𝟐)𝑵 𝒅𝟐𝐞𝐱𝐩[− 𝟏
𝟐𝝈𝟐∑𝒊=𝟏
𝑵 𝒅(𝝆𝑵𝑶 ,𝒅𝒂𝒕𝒂, 𝒊− 𝝆𝑵𝑶 ,𝒔𝒊𝒎𝒖𝒍𝒂𝒕𝒊𝒐𝒏 , 𝒊)𝟐 ]
Time (s)
NO(kg/m
3 )
0.0E+00 5.0E-07 1.0E-06 1.5E-06 2.0E-060.0E+00
5.0E-05
1.0E-04
1.5E-04
2.0E-04
2.5E-04
3.0E-04
Synthetic Data2 Error BarsCandidate Results
MCMC Method - Likelihood
![Page 45: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/45.jpg)
MCMC Method – Metropolis-HastingsEstablish
boundaries for parameter space
Likelihoodcandidate
< Likelihoodcurrent
Likelihoodcandidate
> Likelihoodcurrent
Select initial position
Run simulation at current position
Calculate likelihood for the current
position
Draw new candidate position
Run simulation for candidate position parameters, and
calculate likelihood
Accept or reject candidate
position based on a random number draw
Candidate position is accepted, and becomes
the current chain position
Candidate position becomes
current position
Current position remains
unchanged.
Candidate automatically
accepted
Candidate Accepted
Candidate Rejected
![Page 46: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/46.jpg)
MCMC Method - Improvements
We use the PECOS-developed code QUESO, which implements two major additions to the basic Metropolis-Hastings algorithm. Both of these features can help improve the convergence of the method.• Delayed Rejection:
When an initial candidate position is rejected, a second candidate position is generated based on a scaled proposal covariance matrix.
• Adaptive Metropolis: At periodic intervals, the proposal covariance matrix
is updated based on the calculated covariance of the previously accepted chain positions.
H. Haario, M. Laine, A. Mira, E. Saksman, Statistics and Computing, 16, 339-354 (2006).
![Page 47: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/47.jpg)
Chain Progression
log10(N2 + O <--> NO + N)
log 1
0(O2+N<-->NO+O)
-16.5 -16 -15.5 -15
-18.5
-18
-17.5
-17
log10(N2 + O <--> NO + N)
log 1
0(O2+N<-->NO+O)
-16.5 -16 -15.5 -15
-18.5
-18
-17.5
-17
StartingPosition
log10(N2 + O <--> NO + N)
log 1
0(O2+N<-->NO+O)
-16.5 -16 -15.5 -15
-18.5
-18
-17.5
-17
StartingPosition
CurrentPosition
log10(N2 + O <--> NO + N)
log 1
0(O2+N<-->NO+O)
-16.5 -16 -15.5 -15
-18.5
-18
-17.5
-17
StartingPosition
CurrentPosition
log10(N2 + O <--> NO + N)
log 1
0(O2+N<-->NO+O)
-16.5 -16 -15.5 -15
-18.5
-18
-17.5
-17
StartingPosition
CurrentPosition
log10(N2 + O <--> NO + N)
log 1
0(O2+N<-->NO+O)
-16.5 -16 -15.5 -15
-18.5
-18
-17.5
-17
StartingPosition
CurrentPosition
log10(N2 + O <--> NO + N)
log 1
0(O2+N<-->NO+O)
-16.5 -16 -15.5 -15
-18.5
-18
-17.5
-17
StartingPosition
CurrentPosition
log10(N2 + O <--> NO + N)
log 1
0(O2+N<-->NO+O)
-16.5 -16 -15.5 -15
-18.5
-18
-17.5
-17
StartingPosition
CurrentPosition
log10(N2 + O <--> NO + N)
log 1
0(O2+N<-->NO+O)
-16.5 -16 -15.5 -15
-18.5
-18
-17.5
-17
StartingPosition
Parameter values usedto generate synthetic data
![Page 48: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/48.jpg)
Post-Calibration PDFs
log10(N2 + O <--> NO + N), log10(O2 + N <--> NO + O)
NormalizedPost-CalibrationPDF
-18.5 -18 -17.5 -17 -16.5 -16 -15.5 -150
0.2
0.4
0.6
0.8
1
N2 + O <--> NOO2 + N <--> NO + O
NominalParameterValues
![Page 49: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/49.jpg)
Conclusions
• Global, Monte Carlo based sensitivity analysis can provide a great deal of insight into how various parameters affect a given QoI.• A great deal of computer power is required to perform this type of statistical analysis for DSMC shock simulations.
5,000 shocks were run for this sensitivity analysis. Required a total of ~320,000 CPU hours.
• MCMC can be used to calibrate at least some DSMC parameters based on synthetic data for a 0-D relaxation.• MCMC allows for propagation of uncertainty from the data to the final parameter PDFs.
![Page 50: Bayesian Inference For The Calibration Of DSMC Parameters](https://reader036.vdocuments.net/reader036/viewer/2022062410/5681642c550346895dd5ee9c/html5/thumbnails/50.jpg)
Future Work
• Synthetic data calibrations for a 1-D shock with the current code.• Upgrade the code to allow modeling of ionization and electronic excitation.• Couple the code with a radiation solver.• Sensitivity analysis for a 1-D shock with the additional physics included.• Synthetic data calibrations with the upgraded code.• Calibrations with real data from EAST or similar facility.